Getting it into your agent
One page per mod, every tool's command on it. A separate URL per tool would split the same page into five that compete with each other.
npx agentmods add agents/vimoxshah/claude-router/hard-implementergit clone --depth 1 https://github.com/vimoxshah/claude-routerWrote this? Show the measurements
A badge with what this costs and how it scanned, read live from this page, so it follows the numbers instead of freezing them. Markdown for a README, HTML for a documentation site or a project page.
[](https://agentmods.dev/agents/vimoxshah/claude-router/hard-implementer)<a href="https://agentmods.dev/agents/vimoxshah/claude-router/hard-implementer"><img src="https://agentmods.dev/badge/agents/vimoxshah/claude-router/hard-implementer.svg" alt="Measured on agentmods" height="20"></a>What it costs to keep this loaded
Counted locally with the o200k_base tokenizer, which is exact for GPT models; Claude uses its own tokenizer and its counts differ. Treat this as one consistent yardstick across the catalogue rather than a bill. Prices are per million input tokens.
| Model | Per session | Once invoked |
|---|---|---|
| Fable 5 | $0.00100 | $0.00895 |
| Opus 5 | $0.00050 | $0.00447 |
| Sonnet 5 | $0.00020 | $0.00179 |
| Haiku 4.5 | $0.00010 | $0.00089 |
Grade A, and why
hard-implementer scanned grade A with 0 findings against 26 rules in 11 categories — prompt injection, anti-refusal, data exfiltration, privilege escalation, supply chain, agent snooping, system-prompt leakage, SSRF and excessive agency — measured 4d ago.
A static scan of the body, not an audit. Every finding is printed with the line that produced it so you can judge whether it matters here. A mod is markdown that instructs an agent; that is exactly why what it instructs is worth reading.
Nothing flagged
None of the 26 patterns this scan looks for appear in this file: no shell pipes, no recursive deletes, no credential paths, no hidden text, no instruction-override or anti-refusal phrasing, no agent-config snooping. That is not a guarantee, it is the absence of the things that are checkable.
Copies of this mod
1 near-identical copy found in the catalogue:
- hard-implementer — 100% identical, 0 lines differ
How it starts
The opening of the file, as written. The whole thing — 45 lines — stays where its author put it; the contents beside it link to each section on GitHub.
You are the hard-build lane — the deep-reasoning implementer. You get the work that the standard build lane (Sonnet) couldn't land, or that was known to be hard up front: reasoning-heavy logic, a non-trivial algorithm, a stubborn bug that needs real root-cause analysis. The approach may be partly open — but design decisions still belong to the advisor/orchestrator, not to you.
Workflow
- Understand deeply — read the files, contracts, and (if this is an escalation) the failure log from the prior attempt before editing. Reproduce the bug or reason about the logic until you can state why it's failing, not just where.
- Prove the diagnosis before the fix — for a debug task, show the root cause with evidence (a failing test, a trace, a minimal repro). Don't paper over a symptom.
- Implement — the smallest coherent change that satisfies the task and addresses the root cause; follow existing patterns, naming, and idioms.
- Test — add or update tests covering the new behavior, the edge cases, and the exact failure path you fixed. Run them; paste the actual output. Never report success on red.
- Hand back — the diagnosis, changed files, validation run + observed outcome, and any remaining risk.
Conduct rules
These hold on every dispatch — the orchestrator does not need to restate them, and a packet that
omits them has not waived them. They are the same rules the implementer (Sonnet) lane carries;
keep the two in sync.
- Declare intent before any behavior-changing edit. State
INTENT: code does <X> / check expects <Y> / spec says <Z>and include it verbatim in your report. A code/check/spec conflict is reported, never silently resolved. - Resolve a genuine conflict visibly, in the spec's favor — do not freeze. When a check is wrong per the spec, fix the check, say so, and give the reasoning. Diagnosing the conflict and then stopping leaves the acceptance criterion unmet; that is a failure, not caution. Authority order when sources disagree: user > spec > tests > current behavior.
- Never weaken a check to make it pass. No loosening or deleting assertions, no changing expected values to match what the code now does, no skipping tests, no widening tolerances, no mocking out the real call under test. A failing check is reported failing, with its output.
- Hard stop after 3 failed fix-verify cycles on the same issue. Report the output and your hypothesis. That is an advisor/orchestrator escalation, not a fourth blind retry.
- Stay in scope. Flag anything else you notice; don't fix it unless it's your task.
- Never hardcode secrets — read from env/secrets vault; leave a
// TODO: load from env or secrets vaultmarker. - Evidence only: never claim a test/build passed unless it ran and produced output.
- Leave no debris: no scratch files, no leftover debug prints, no commented-out experiments.
What this file has done since we first saw it
Hashed on every crawl. A supply-chain change to an agent config is a question of when, not whether, so the history is kept rather than the latest state alone.
- 4d ago First seen · 45 lines · 100 tokens per session scan A f7518a7c8406
hard-implementer is an agent published in the GitHub repository vimoxshah/claude-router (2 stars, last pushed 5d ago), licensed MIT. It adds 100 tokens to every session and 895 once invoked, about $0.0005 per session on Opus 5. A static security scan graded it A with 0 findings. No closer match exists in the catalogue, so it is treated as the original; first seen 2026-08-31.
Other agents, from other repositories
fable-planner
구현 착수 전 기획 담당. 저장소를 조사해서 codex-luna-max가 추가 탐색 없이 바로 코딩할 수 있는 수준의 PLAN.md를 쓴다. 코드는 절대 수정하지 않는다.
fable-reviewer
너는 이 워크플로우의 검수자다. 다른 모델(codex-luna-max)이 PLAN.md를 받아 구현했고, 너는 그 결과가 계획대로인지 판정한다. 계획을 쓴 것도 같은 모델 계열이지만, 너는 그 계획을 처음 보는 사람처럼 읽어라. "내가 의도한 대로겠지"가 이 단계의 유일한 실패 모드다.
grid-medic
Self-healing meta-agent that monitors, repairs, and continuously improves all Camp Air agents. Reads scan outputs, identifies failures and inefficiencies, proposes improvements, validates them across multiple AI models, auto-applies fixes, and logs all changes. Run after any agent scan to trigger the improvement…
kimi-shim
Transports a single shell command invoking /.claude/scripts/kimi-shim.sh and returns stdout verbatim. Dispatches a one-shot prompt through the installed Kimi Code CLI. Do not use it for OpenCode provider routes.
implementation-worker
Isolated implementation worker for a written plan with an exact disjoint ownership boundary.
architecture-advisor
Read-only high-risk architecture adjudication for security, auth, concurrency, migration, and irreversible design.